On June 1, 2026, MiniMax announced MiniMax M3, which the company calls the first and only open-weight model to combine three traits that had previously lived in separate systems: frontier-level coding, a 1-million-token context window, and native multimodality across image and video input. The June 1, 2026 date is stated in the announcement header on the official MiniMax blog. MiniMax said it would release the model’s technical report and open-source the corresponding model weights within 10 days of the announcement (claim traces to the primary).
The central architectural change is MSA (MiniMax Sparse Attention), a proprietary attention mechanism that supports context windows of up to 1 million tokens. At maximum context length, MiniMax states M3’s per-token compute is just 1/20 that of its previous-generation model (claim traces to the primary). M3 is described as a natively multimodal model that supports image and video input and can operate a desktop computer, with mixed-modality training from the start. On reported benchmarks it scores 59.0% on SWE-Bench Pro, 66.0% on Terminal-Bench 2.1, and 74.2% on MCP Atlas (all numbers traced to the primary). The M3 API went live the same day, with pricing tiered by input length.
For business and technical leaders, M3 matters because it pushes open-weight models further into territory that had been dominated by closed frontier labs, specifically long-context agentic coding and multimodal computer use. An open-weight option at this capability level changes the build-versus-buy calculus for teams that want to self-host or fine-tune rather than depend on a closed API.